We show that commonly used part-of-speech (POS) taggers, despite their high reported performance, in many cases make tagging errors on simple and unambiguous sentences. We collect a new data set of non-ambiguous sentences that can easily be tagged by human taggers, but where at least one standard POS tagger makes precisely one tagging error. Furthermore, we present a method for generating rules that are meant to correct the output of a standard POS tagger. Applying this method to the new data set, we extract a set of such rules, which are then evaluated over another data set introduced in earlier work. Our results show that the method works, but also that the increase in tagging accuracy is rather small, probably due to the small size of our training data set. Finally, we present an analysis of POS tagging in general, concluding that there are multiple ambiguities that introduce unresolvable challenges in POS tagging.